{
  "note": "Registry of result labels. 'order' of this list fixes the bootstrap RNG stream: append new models at the end so existing confidence intervals never move.",
  "models": [
    {
      "label": "alibaba/qwen-flash",
      "model_ref": "qwen-flash",
      "route": "Alibaba Cloud Model Studio API, buffered (non-streaming), enable_thinking=false, temperature 0",
      "hardware_class": "hosted API; serving hardware not disclosed",
      "output_token_cap": {"R1-R5": 512, "R6-R7": 2048},
      "latency_instrument": "Buffered API: wall time only; TTFT and tok/s unavailable"
    },
    {
      "label": "alibaba/qwen-plus",
      "model_ref": "qwen-plus",
      "route": "Alibaba Cloud Model Studio API, buffered (non-streaming), enable_thinking=false, temperature 0",
      "hardware_class": "hosted API; serving hardware not disclosed",
      "output_token_cap": {"R1-R5": 512, "R6-R7": 2048},
      "latency_instrument": "Buffered API: wall time only; TTFT and tok/s unavailable",
      "coverage_note": "Run stopped during R6; R6 partial and R7 not run. Missing trials are not failures."
    },
    {
      "label": "RK1-Qwen3-4B",
      "model_ref": "Qwen3-4B-Instruct-2507, RKLLM W8A8 build, compiled context 2048",
      "route": "owned edge node, local NPU inference via the RKLLM runtime (1.3.1)",
      "hardware_class": "edge SoC: Turing RK1 compute module, Rockchip RK3588, 3-core NPU, 32 GB",
      "output_token_cap": {"R1-R5": 512, "R6-R7": 512},
      "latency_instrument": "RKLLM native callbacks/perf; trial wall time includes remote shell and model init",
      "coverage_note": "Native 512 output / 2048 context limit kept for R6-R7, unlike the 2048 output cap on hosted routes. Input tokens were not logged."
    },
    {
      "label": "astra",
      "model_ref": "openai/gpt-6-astra",
      "route": "OpenRouter, OpenAI-compatible streaming chat completions, temperature 0",
      "hardware_class": "hosted API; serving hardware not disclosed",
      "output_token_cap": {"R1-R5": 512, "R6-R7": 2048},
      "latency_instrument": "SSE first content event; tok/s = (completion_tokens-1)/(last-first content event)"
    },
    {
      "label": "deepseek-v4-flash",
      "model_ref": "deepseek/deepseek-v4-flash",
      "route": "OpenRouter, OpenAI-compatible streaming chat completions, temperature 0",
      "hardware_class": "hosted API; serving hardware not disclosed",
      "output_token_cap": {"R1-R5": 512, "R6-R7": 2048},
      "latency_instrument": "SSE first content event; tok/s = (completion_tokens-1)/(last-first content event)"
    },
    {
      "label": "glm-4.7-flash",
      "model_ref": "z-ai/glm-4.7-flash",
      "route": "OpenRouter, OpenAI-compatible streaming chat completions, temperature 0",
      "hardware_class": "hosted API; serving hardware not disclosed",
      "output_token_cap": {"R1-R5": 512, "R6-R7": 2048},
      "latency_instrument": "SSE first content event; tok/s = (completion_tokens-1)/(last-first content event)"
    },
    {
      "label": "gpt-5-nano",
      "model_ref": "openai/gpt-5-nano",
      "route": "OpenRouter, OpenAI-compatible streaming chat completions, temperature 0",
      "hardware_class": "hosted API; serving hardware not disclosed",
      "output_token_cap": {"R1-R5": 512, "R6-R7": 2048},
      "latency_instrument": "SSE first content event; tok/s = (completion_tokens-1)/(last-first content event)"
    },
    {
      "label": "qwen3.7-flash",
      "model_ref": "qwen/qwen3.7-flash",
      "route": "OpenRouter, OpenAI-compatible streaming chat completions, temperature 0",
      "hardware_class": "hosted API; serving hardware not disclosed",
      "output_token_cap": {"R1-R5": 512, "R6-R7": 2048},
      "latency_instrument": "SSE first content event; tok/s = (completion_tokens-1)/(last-first content event)"
    }
  ]
}